From mboxrd@z Thu Jan 1 00:00:00 1970 From: Avi Kivity Subject: Re: [PATCH 1/4] KVM: Using kfifo for irq recording Date: Thu, 25 Dec 2008 15:26:29 +0200 Message-ID: <49538A05.1020104@redhat.com> References: <1230019973-16833-1-git-send-email-sheng@linux.intel.com> <1230019973-16833-2-git-send-email-sheng@linux.intel.com> <4953696A.5070307@redhat.com> <200812251927.39118.sheng@linux.intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: kvm@vger.kernel.org To: Sheng Yang Return-path: Received: from mx2.redhat.com ([66.187.237.31]:38076 "EHLO mx2.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750927AbYLYN0d (ORCPT ); Thu, 25 Dec 2008 08:26:33 -0500 In-Reply-To: <200812251927.39118.sheng@linux.intel.com> Sender: kvm-owner@vger.kernel.org List-ID: Sheng Yang wrote: > On Thursday 25 December 2008 19:07:22 Avi Kivity wrote: > >> Sheng Yang wrote: >> >>> For MSI-X, we have to deal with multiply IRQ with same IRQ handler, so >>> it's necessary to record the IRQ that trigger the IRQ handler. >>> >> Does MSI-X disallowing coalescing two requests into one interrupt? Or >> can we still coalesce interrupts (perhaps by recording them as a (irq, >> cpu) pair?) >> > > Disallow? Not quite understand. PCI spec said OS don't need to ensure the > sequence they handled is the same as they happened. This struct is used just > because we lost information of irq after schedule_work... > > Why can't we store this information in a bitmap? There are a limited number of irqs. The only reason I can think of for using a fifo is if we want to preserve the number and ordering of interrupts. Is there another reason? >>> @@ -313,6 +314,9 @@ struct kvm_assigned_dev_kernel { >>> int host_irq; >>> bool host_irq_disabled; >>> int guest_irq; >>> +#define KVM_ASSIGNED_DEV_IRQ_FIFO_LEN 0x100 >>> + struct kfifo *irq_fifo; >>> + spinlock_t irq_fifo_lock; >>> #define KVM_ASSIGNED_DEV_GUEST_INTX (1 << 0) >>> >> What if it runs out? >> >> What does real hardware do? I'm sure it doesn't have a 100-entry queue. >> > > 0x100 is just a simple number which I thought different interrupts of same > MSI-X device can happen at same period(indeed it's 0x100/sizeof(int)). Maybe > not that many. And it just used by work function later to find what guest > vector is, and then inject the correlated interrupt to the guest. > Maybe it's better to do the conversion immediately, so we can store the information in a structure that's not prone to overflow. -- error compiling committee.c: too many arguments to function